Roboticist Georgia Chalathi: scale is learning to use structure, not replacing it
GeorgiaChal · x · 2026-09-22
Georgia Chalathi expands on the "agentic robotics" paradigm discussion with Ken Goldberg, where she questioned how much robot performance comes from the model vs the harness. Her argument:
- Scale isn't making structure obsolete—scale is becoming capable of using structure.
- Two levels of structure: around the model (action spaces, tools, controllers, simulators, safety mechanisms) and inside learned representations (equivariance, object/interaction representations, contact and physical constraints).
- Increasingly capable models like Astra can select, compose and even construct structure, not just consume it.
The question is no longer structure vs scale, but where structure should live: what models learn themselves vs what stays external.
More from AGI Musings
- Insects-matter-more-than-humans debate sparks 'moral realism fallacy' rejoinder — mjdramstead · 2026-09-22
- Bentham's Bulldog doubles down on insect-worth article as critics demand premise rejection — mjdramstead · 2026-09-22
- Smaller models aren't always cheaper: the hidden costs LLM business cases miss — zeuslac · 2026-09-22
- Mathematicians get special treatment from AI labs while entry-level SWE market gets crushed, critics argue — eigenrobot · 2026-09-22
- Debate: will all mainstream software be formally verified by 2030? Legacy debt says no — luislamb · 2026-09-22
- Alibaba CEO Eddie Wu: Machine thinking is under 3% of human thinking, with 1000x runway ahead — alifcoder · 2026-09-22